A gear residual service life prediction method based on deep adaptive fusion

Abstract

This invention discloses a method for predicting the remaining service life of gears based on deep adaptive fusion, belonging to the field of rotating machinery fault diagnosis technology. It includes an offline stage and an online stage. In the offline stage, an accelerated life test is performed on the gear to be predicted, and the gear's acceleration signal is collected. A health index corresponding to the gear to be predicted is established. The health index corresponds to the time-frequency domain features of the acceleration signal of the gear to be predicted under different remaining service lives. In the online stage, the acceleration signal of the gear to be predicted is collected, and the fused signal features are extracted. The fused signal features are then regressed and mapped to the health index to obtain the remaining service life of the gear to be predicted. The beneficial effects of this invention are: accurate extraction and fusion of dynamic multi-scale long-range dependency information in vibration signals, achieving high-precision prediction of remaining service life, and providing theoretical support for gear health assessment.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and specifically to a method for predicting the remaining service life of gears based on deep adaptive fusion. Background Technology

[0002] As the core power transmission component of a mechanical transmission system, gears inevitably endure the coupling effects of multiple physical fields, such as friction and wear, cyclic impact loads, and alternating stress, during continuous service. This causes a gradual decline in key service performance characteristics, such as tooth surface contact fatigue strength and meshing accuracy, ultimately leading to failures such as pitting and tooth breakage. Therefore, accurately predicting the remaining service life of gears is of great significance for ensuring the safety and reliability of equipment and extending its service life. Existing technologies include methods for predicting gear life by reading characteristics such as gear vibration and the sounds emitted during operation.

[0003] For example, patent application CN202410557309.4 discloses a technology in the field of fault diagnosis and health management of rotating machinery, solving the technical problem that traditional methods cannot quickly and accurately predict the remaining service life of gearboxes. Specifically, it relates to a gearbox life prediction method based on BiSACvLSTM. This method includes the following steps: acquiring the vibration signal of gearbox acceleration, and decomposing the vibration signal into several Intrinsic Mode Functions (IMFs) and residuals using EEMD; selecting IMFs with more degradation characteristics from the IMFs using the similarity metric EISI; and constructing a BiSACvLSTM network for predicting the remaining service life (RUL) of the gearbox. This prediction method can quickly and accurately predict the remaining service life of gearboxes with low computational complexity, while improving the accuracy and robustness of gearbox RUL prediction.

[0004] For example, patent application CN201910432767.4 discloses a method for predicting the real-time remaining life of gears by monitoring multiple degradation quantities, belonging to the field of mechanical reliability technology. The implementation steps are as follows: 1. Real-time monitoring of gear degradation in the main test gearbox using accelerometers and noise sensors; 2. Feature extraction and degradation assessment of the gear degradation state; 3. Modeling the gearbox vibration acceleration and noise using kernel estimation and stochastic filtering theory to obtain the gearbox remaining life probability density function and the marginal distribution function of the remaining life for a single degradation quantity; 4. Using the Copula function to represent the stochastic correlation between the gearbox vibration acceleration and noise, obtaining the joint distribution function of the gearbox remaining life; 5. Obtaining the joint probability density function of the remaining life based on the joint distribution function of the gearbox remaining life, and finally obtaining the predicted value of the gearbox remaining life. The advantage is that it effectively predicts the gear degradation state and real-time remaining life, providing a basis for preventive maintenance of gears.

[0005] However, in actual implementation, the inventors found that this type of technical solution has limited ability to extract local features of vibration signals, and the fixed time-frequency resolution makes it difficult to fully reveal the degradation process of gears, resulting in insufficient accuracy and robustness in the prediction of remaining service life. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, a method for predicting the remaining service life of gears based on deep adaptive fusion is provided.

[0007] The specific technical solution is as follows: A method for predicting the remaining service life of gears based on deep adaptive fusion includes an offline stage and an online stage; in the offline stage, an accelerated life experiment is performed on the gear to be predicted and the acceleration signal of the gear is collected, and a health index corresponding to the gear to be predicted is established; the health index corresponds to the time-frequency domain features of the acceleration signal of the gear to be predicted under different remaining service lives; in the online stage, the acceleration signal of the gear to be predicted is collected and the fused signal features are extracted, and the fused signal features are regressed and mapped to the health index to obtain the remaining service life of the gear to be predicted.

[0008] On the other hand, the offline stage includes: Step A1: conducting an accelerated life test on the gear to be predicted until the gear to be predicted is damaged, collecting the acceleration signal as experimental time-domain data during the experiment, and performing a fast Fourier transform on the experimental time-domain data to obtain experimental frequency-domain data; Step A2: estimating the probability distribution of the experimental time-domain data and the experimental frequency-domain data respectively, and calculating the time-domain health index and the frequency-domain health index using the intersection-union ratio of the distributions; Step A3: calculating the index weights of the time-domain health index and the frequency-domain health index respectively, and constructing the health index.

[0009] On the other hand, in step A2, the probability distribution of the experimental time-domain data is fitted based on a Gaussian mixture model, and the probability distribution of the experimental frequency-domain data after power transformation is estimated based on a Box-Cox mixture model.

[0010] On the other hand, the indicator weights in step A3 include:

[0011] ;

[0012] In the formula, For the current time domain weights, The time-domain weights of the previous period, For the current frequency domain weights, The frequency domain weights of the previous period;

[0013] For smoothing coefficients, For the time-domain health indicators, This refers to the frequency domain health index.

[0014] On the other hand, the online stage includes: Step B1: Acquiring the acceleration signal of the gear to be predicted and decomposing it through a high-pass filter and a low-pass filter of multi-level discrete wavelet transform to obtain detail coefficients and approximate coefficients, and assembling them to obtain a coefficient sequence; Step B2: Dividing and embedding the coefficient sequence to obtain an embedded coefficient sequence; Step B3: Extracting local features and global features from the embedded coefficient sequence respectively, and fusing them to obtain fused signal features; Step B4: Predicting the remaining service life based on the fused signal features.

[0015] On the other hand, step B2 includes: step B21: splitting the coefficient sequence into multiple frame blocks according to a preset length; step B22: projecting the frame blocks onto a feature space of a specific dimension to obtain the embedded coefficient sequence.

[0016] On the other hand, step B3 includes: step B31: extracting wavelet inputs at different decomposition levels from the embedded coefficient sequence and inputting them into an LSTM model to obtain local features, and inputting the embedded coefficient sequence into a Transformer model to extract global features; step B32: fusing the local features and the global features to obtain the fused signal features.

[0017] On the other hand, step B32 includes: calculating the weights of the local features and the global features respectively based on the self-attention mechanism and then fusing them.

[0018] On the other hand, step B4 includes: step B41: using a multilayer perceptron to perform nonlinear mapping based on the fused signal characteristics to obtain the corresponding health index; step B42: determining the remaining service life based on the health index.

[0019] A storage medium includes computer instructions that, when executed by a computer device, perform the aforementioned method for predicting the remaining service life of a gear.

[0020] The above technical solution has the following advantages or beneficial effects: Addressing the problem that existing technologies are insufficient in extracting time-frequency domain information from gear vibration signals and struggle to reflect the characteristics of strongly time-varying, non-stationary vibration signals, a Global-Local Depth Adaptive Fusion (GLDAF) model based on multi-scale decomposition is established. This model captures local features of vibration signals from multiple scales, compensating for the shortcomings of existing methods in recognizing local features when processing non-stationary signals. It can accurately extract and fuse long-range dependency information of dynamic multi-scale vibration signals, achieving high-precision prediction of remaining service life and providing theoretical support for gear health assessment. Attached Figure Description

[0021] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0022] Figure 1 This is an overall schematic diagram of an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of the offline stage in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of the online stage in an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of step B2 in an embodiment of the present invention;

[0026] Figure 5 This is a schematic diagram of step B3 in an embodiment of the present invention;

[0027] Figure 6 This is a schematic diagram of step B4 in an embodiment of the present invention;

[0028] Figure 7 This is a schematic diagram illustrating the construction results of gear health indicators in an embodiment of the present invention;

[0029] Figure 8 This is a schematic diagram illustrating the prediction results of the remaining service life in an embodiment of the present invention;

[0030] Figure 9This is a schematic diagram showing the uncertainty quantification estimation results of the sampling points at the 60th and 120th minutes in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0034] This invention includes: a method for predicting the remaining service life of gears based on deep adaptive fusion, such as... Figure 1 As shown, it includes an offline stage and an online stage; in the offline stage, an accelerated life test is performed on the gear to be predicted and the acceleration signal of the gear is collected, and a health index corresponding to the gear to be predicted is established.

[0035] The health index corresponds to the time-frequency domain features of the acceleration signal of the gear to be predicted under different remaining service lives. In the online stage, the acceleration signal of the gear to be predicted is collected and the fused signal features are extracted. The fused signal features are regressed and mapped to the health index to obtain the remaining service life of the gear to be predicted.

[0036] Specifically, since the construction of health indicators plays a decisive role in the accuracy and robustness of remaining service life prediction, this method first constructs a new gear health status index based on an adaptive weighted cross-domain hybrid model (AW-CDMM). Addressing the problem that existing remaining service life predictions struggle to reflect the characteristics of strongly time-varying, non-stationary vibration signals, a Global-Local Deep Adaptive Fusion (GLDAF) model based on multi-scale decomposition is established. This model captures local features of vibration signals from multiple scales, overcoming the shortcomings of existing methods in recognizing local features when processing non-stationary signals. By combining deep learning and adaptive fusion strategies, this invention can accurately extract and fuse long-range dependency information from dynamic multi-scale vibration signals, achieving high-precision prediction of remaining service life and providing theoretical support for gear health assessment.

[0037] Specifically, the above-mentioned solution is implemented in two main parts: an offline phase and an online phase. The offline phase refers to the process of conducting accelerated life tests on a specific gear to be predicted, such as the gearbox main shaft gear, on an accelerated life test bench, and using vibration signal sensors, multi-axis accelerometers, etc., to acquire vibration data reflecting gear degradation information to calibrate the corresponding health indicators. This process can determine the time-frequency domain characteristics of the signal under different remaining service life stages through the health indicators.

[0038] The online phase refers to the process where the gear to be evaluated has been installed in the corresponding product. For the gear to be evaluated that needs to be monitored, long-term vibration signals are collected using a multi-axis accelerometer. Then, a segment is extracted from the acceleration signal sequence, local features of the vibration signal are captured from multiple scales, and the long-range dependence information of dynamic multi-scale in the vibration signal is fused. Then, the regression mapping is applied to health indicators to accurately predict the remaining service life of the gear to be evaluated.

[0039] In one embodiment, such as Figure 2 As shown, the offline stage includes: Step A1: Conduct an accelerated life test on the gear to be predicted until the gear to be predicted is damaged. During the experiment, the acceleration signal is collected as the experimental time-domain data, and the experimental time-domain data is subjected to a fast Fourier transform to obtain the experimental frequency-domain data; Step A2: Estimate the probability distribution of the experimental time-domain data and the experimental frequency-domain data respectively, and use the intersection-union ratio of the distributions to calculate the time-domain health index and the frequency-domain health index; Step A3: Calculate the index weights of the time-domain health index and the frequency-domain health index respectively, and construct the health index.

[0040] Specifically, to achieve better calibration results, this embodiment introduces both time-domain and frequency-domain features during the construction of health indicators to better characterize the vibration signal features at different life stages. This includes first acquiring gear acceleration signals using an accelerated life testing platform to obtain vibration data reflecting gear degradation as experimental time-domain data, and then performing a fast Fourier transform on the experimental time-domain data to obtain experimental frequency-domain data.

[0041] Subsequently, corresponding processing steps are performed for the time-frequency domain characteristics, including fitting the probability distribution of time-domain data based on a Gaussian mixture model, estimating the probability distribution of time-domain data throughout the entire life cycle, and calculating the time-domain health index of the gear using the intersection-union ratio of the distributions; performing a Fourier transform on the original time-domain data to obtain the frequency domain signal of vibration acceleration; estimating the probability distribution of the power-transformed vibration frequency domain signal based on a Box-Cox mixture model, estimating the probability distribution of frequency domain data throughout the entire life cycle, and calculating the frequency domain health index of the gear using the intersection-union ratio of the distributions.

[0042] Finally, to improve the recognizability of features, the adaptive filtering of AW-CDMM is used to dynamically adjust the weights of time-domain and frequency-domain health indicators, and weighted fusion is performed to ultimately construct the gear health indicator.

[0043] In one embodiment, in step A2, the probability distribution of the experimental time-domain data is fitted based on a Gaussian mixture model, and the probability distribution of the experimental frequency-domain data after power transformation is estimated based on a Box-Cox mixture model.

[0044] Specifically, to achieve better processing results, in this embodiment, an accelerometer is first installed on the gear accelerated life test bench to collect the vibration signal of the gear as experimental time-domain data. Then, it is converted into an experimental frequency domain signal using Fourier transform. X =( x 1, x 2, ..., x M ), x =( x 1, x 2, ..., x m ).

[0045] Then, the probability distribution of the experimental time-domain data was fitted based on a Gaussian mixture model, the probability distribution of the experimental frequency-domain data after power transformation was estimated based on a Box-Cox mixture model, and the time-domain health index was calculated using the intersection-union ratio of the distributions. and frequency domain health indicators .

[0046] The specific steps for calculating the intersection-union ratio (IU) include constructing an AW-CDMM system to build health indicators, including estimating the baseline distribution using the first L healthy samples, and including the time-domain baseline distribution of health status. Frequency domain baseline distribution of health status Furthermore, the entire lifecycle samples are estimated across domains using AW-CDMM, including the temporal distribution of the entire lifecycle. And the time-domain and frequency-domain X of the entire life cycle, respectively and P = ( P 1, P 2, ..., P M Based on the baseline distribution of health data and the distribution of life-cycle data, the health index of the gear is calculated through the intersection-union ratio of the distributions. The health index equation is expressed as:

[0047]

[0048] in, H tand H f These are time-domain health indicators and frequency-domain health indicators, ω t and ω f These represent time-domain weights and frequency-domain weights, respectively, indicating the degree of influence of time-domain data and frequency-domain data on health indicators. This describes the calculation process for the intersection-union ratio of the distributions. Time-domain baseline distribution corresponding to health status , Corresponding to the actual characteristics of time-domain data, The frequency domain baseline distribution corresponding to the health state, This corresponds to the actual frequency domain data characteristics.

[0049] Finally, the health indicators of the gear degradation process are: HI = ( H 1, H 2, ..., H M This comprehensively indicates the health status and degradation trend of gears across different domains.

[0050] In one embodiment, the indicator weights in step A3 include:

[0051] ;

[0052] In the formula, For the current time domain weights, The time domain weights of the previous period, For the current frequency domain weights, The frequency domain weights of the previous cycle; For smoothing coefficients, For time-domain health indicators, This is a frequency domain health indicator.

[0053] Specifically, to achieve better recognition accuracy, this embodiment constructs the aforementioned adaptive filtering function to generate each health indicator. H M The corresponding fusion weights. Among them, HI = ( H 1, H 2, ..., H M The health indicators corresponding to different periods M are generated sequentially. Based on the above adaptive filter, dynamically updated fusion weights are generated to adapt to the changes in the significance of vibration signals in the time and frequency domain at different life stages.

[0054] In one embodiment, such as Figure 3As shown, the online stage includes: Step B1: Acquire acceleration signals of the gear to be predicted and decompose them using high-pass and low-pass filters of multi-level discrete wavelet transform to obtain detail coefficients and approximate coefficients, and assemble them to obtain a coefficient sequence; Step B2: Divide and embed the coefficient sequence to obtain an embedded coefficient sequence; Step B3: Extract local features and global features from the embedded coefficient sequence respectively, and fuse them to obtain fused signal features; Step B4: Predict the remaining service life based on the fused signal features.

[0055] Specifically, to achieve better regression prediction, in this embodiment, during the online detection phase, acceleration signals of the gear to be predicted are first collected using a multi-axis accelerometer. To eliminate differences in measurement equipment, reversible instance normalization processing is typically required after data acquisition.

[0056] Since the signal sequence in online measurement is long, multi-scale processing is required. This includes using high-pass and low-pass filters based on multi-level discrete wavelet transform to decompose the signal into detail coefficients and approximate coefficients to form a coefficient sequence. The detail coefficients and approximate coefficients are then divided into blocks and embedded to obtain an embedded coefficient sequence.

[0057] Then, the coefficient sequences after multi-scale decomposition are input into separate LSTM modules for local feature extraction. At the same time, the coefficient sequences after all blocks are embedded are concatenated and input into the Transformer module for global feature extraction. Based on learnable attention weights, adaptive fusion of global and local features is achieved. The fused features are nonlinearly mapped through a multilayer perceptron to achieve the regression output of the remaining service life of the gear.

[0058] In one embodiment, such as Figure 4 As shown, step B2 includes: step B21: splitting the coefficient sequence into multiple frame blocks according to a preset length; step B22: projecting the frame blocks onto a feature space of a specific dimension to obtain an embedded coefficient sequence.

[0059] Specifically, to achieve better feature representation, in this embodiment, the normalized acceleration signal is decomposed using a high-pass filter and a low-pass filter of multi-level discrete wavelet transform, including:

[0060] ;

[0061] in, X Ai and X Di The first i Approximation coefficients and detail coefficients at each decomposition level ζ It is a wavelet type. m To decompose the series, This represents the transpose of the normalized acceleration signal.

[0062] In this process, the high-pass filter outputs high-frequency components, i.e., detail coefficients, while the low-pass filter outputs low-frequency information, i.e., approximation coefficients. Furthermore, since the approximation coefficients before the final level will be further decomposed into new detail coefficients and approximation coefficients, to avoid information redundancy, only the approximation coefficients of the final level are retained here, while all detail coefficients from all levels are preserved.

[0063] The approximation coefficients and detail coefficients are then assembled into a coefficient sequence. ;

[0064] In the formula, For the coefficient sequence, Let be the approximation coefficient of the m-th term. Let be the detail coefficient of the m-th term.

[0065] Then, to facilitate subsequent model extraction, the assembled coefficient sequence is cut into frame blocks of length P, which can overlap.

[0066] The block division result is:

[0067] ;

[0068] In the formula, For the block results, For the coefficient sequence, N i The corresponding total number of frames after the blocks are divided. N i = [( L i - P ) / S ] + 2, L i The total length corresponding to the i-th coefficient sequence, S is the step size of the non-overlapping part.

[0069] Then, the embedding process is performed:

[0070] ;

[0071] It is the embedding coefficient sequence.

[0072] In one embodiment, such as Figure 5 As shown, step B3 includes: step B31: extracting wavelet inputs of different decomposition levels from the embedded coefficient sequence and inputting them into the LSTM model to obtain local features, and inputting the embedded coefficient sequence into the Transformer model to extract global features; step B32: fusing the local features and global features to obtain fused signal features.

[0073] Specifically, to achieve feature extraction after embedding, in this embodiment, for local features, wavelet inputs at different decomposition levels are adjusted and fed into the pre-trained LSTM model to extract local features at different scales while preserving the correlation between input signals at different time series. Simultaneously, for the global signal, the embedding coefficient sequence is directly input into the Transformer model to extract global features.

[0074] Subsequently, the local and global features are fused to obtain the fused signal features.

[0075] In one embodiment, step B32 includes: calculating the weights of local features and global features respectively based on a self-attention mechanism and then fusing them.

[0076] Specifically, to achieve better feature fusion results, this embodiment introduces a self-attention mechanism for fusion, thereby calculating the weight coefficients of each path:

[0077] ;

[0078] ;

[0079] a The attention weights are for the corresponding features. W a and b a As learnable parameters, Softmax guarantees that the sum of all weights is 1. H fuse The global features output after fusion. H i For the first i Each frame block feature. The fusion matrix is ​​obtained by concatenating local and global features:

[0080] H cat = [H LSTM H Trans ], where H LSTM For local features, H Trans This is a global feature.

[0081] In one embodiment, such as Figure 6 As shown, step B4 includes: step B41: using a multilayer perceptron to perform nonlinear mapping based on the fused signal characteristics to obtain the corresponding health indicators; step B42: determining the remaining service life based on the health indicators.

[0082] Specifically, to achieve better regression results, after establishing the fused signal features, a multilayer perceptron is used for nonlinear mapping:

[0083] ;

[0084] in, For activation function, W 1 and W 2 represents the weight. b 1 and b 2 is the bias parameter.

[0085] This mapping process can match signals with pre-established health indicators and obtain the corresponding remaining lifespan.

[0086] In the actual training process, error evaluation metrics such as MAE, RMSE, MAPE, and Score are also used for comparison to determine the accuracy of the method. Based on AKDE, the error of the method's prediction results is estimated in intervals, thereby performing an uncertainty quantification evaluation of the prediction method's performance.

[0087] Specifically, it includes:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] in, ;

[0093] In the formula, , y s , These are the true value and the predicted value, respectively.

[0094] For AKDE validation, assuming the prediction error dataset is... P =[ p 1, p 2, …, p n ],in n Given the number of sample points, the AKDE verification is as follows:

[0095] ;

[0096] in, p i For the first i One prediction error value, s The number of dimensions of the variable. K This is the Gaussian kernel function. k h The bandwidth ish i The kernel density estimation function, ω For bandwidth h i The weight, , α The sensitivity factor is usually set to 0.5. α When the value is 0, the adaptive bandwidth becomes the fixed bandwidth.

[0097] The above method was tested, and three sets of gear lifecycle datasets were collected under different operating conditions, as shown in Table 1. Experimental results show that the total lifespan of the tested gearbox under the three operating conditions was 16524 min, 13419 min, and 14106 min, respectively. Gear failures were concentrated on tooth surface wear, pitting, and tooth breakage in the second-stage driving gear and the second-stage driven gear. Therefore, the vibration acceleration signal of the second-stage driving gear during the last 1200 minutes of degradation was selected to study the remaining service life of the gear.

[0098] Table 1

[0099]

[0100] Based on the three sets of vibration datasets mentioned above, the AW-CDMM was used to construct the health index HI, such as... Figure 7 As shown, the HI constructed by this method exhibits a monotonically decreasing trend, effectively characterizing the degradation trend of gears over time. In the initial stage, HI is close to 1, indicating that the current state of the gear differs little from the distribution of healthy data, i.e., it is in a healthy state. As the running time increases, HI gradually decreases, indicating that the distribution difference between the current state and healthy data becomes larger, leading to different types of faults on the gear surface. Finally, when HI reaches a certain value, the gear will experience serious faults and cease to function properly.

[0101] The GLDAF prediction method based on multi-scale decomposition was used to predict the remaining lifetime of three datasets. First, all HI data were divided into training and test sets in a 9:1 ratio, with 90% of the data used for model training and 10% for model evaluation. Specifically, the remaining lifetime of the last 120 minutes was predicted using the first 1080 minutes of HI data. The decomposition scale was set to 3 levels. During model training, the Adam optimizer was selected with a learning rate of 0.0001, and Dropout regularization (with a Dropout value of 0.2) was used to prevent overfitting. Figure 8 The actual, training, and predicted values ​​of HI for three datasets are described, with the shaded area representing the 95% confidence interval. The error between the training and actual HI values ​​is small, the predicted values ​​show the same trend as the actual values, and all predicted values ​​are within the 95% confidence interval.

[0102] To further illustrate the accuracy of the GLDAF method, the MAE, RMSE, MAPE, and Score between the predicted and actual values ​​were calculated, as shown in Table 2. The results show that the method proposed in this invention has high prediction accuracy and small fluctuations in the prediction error.

[0103] Table 2

[0104]

[0105] To further and more intuitively evaluate the accuracy of the HI interval prediction for the health indicator, interval prediction estimates were performed at the 60-minute and 120-minute intervals to be predicted. Figure 9 As shown, the true values ​​of HI are all within the 95% confidence interval, and the actual values ​​basically coincide with the estimated values, indicating the effectiveness and reliability of the proposed method, which can provide strong support for risk assessment and decision-making regarding equipment health status.

[0106] A storage medium includes computer instructions that, when executed by a computer device, perform the aforementioned method for predicting the remaining useful life of a gear.

[0107] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of gears based on deep adaptive fusion, characterized in that, Includes offline and online phases; During the offline phase, an accelerated life test is performed on the gear to be predicted and the acceleration signal of the gear is collected, and a health index corresponding to the gear to be predicted is established. The health index corresponds to the time-frequency domain characteristics of the acceleration signal of the gear to be predicted under different remaining service lives; During the online phase, the acceleration signal of the gear to be predicted is collected and the fused signal features are extracted. The fused signal features are then regressed and mapped to the health index to obtain the remaining service life of the gear to be predicted. The offline phase includes: Step A1: Perform an accelerated life test on the gear to be predicted until the gear to be predicted is damaged. During the test, collect the acceleration signal as experimental time-domain data, and perform a fast Fourier transform on the experimental time-domain data to obtain experimental frequency-domain data. Step A2: Estimate the probability distribution of the experimental time-domain data and the experimental frequency-domain data respectively, and calculate the time-domain health index and the frequency-domain health index using the intersection-union ratio of the distributions; Step A3: Calculate the weights of the time-domain health index and the frequency-domain health index respectively, and construct the health index; The indicator weights in step A3 include: ； In the formula, For the current time domain weights, The time-domain weights of the previous period, For the current frequency domain weights, The frequency domain weights of the previous period; For smoothing coefficients, For the time-domain health indicators, This refers to the frequency domain health index.

2. The method for predicting the remaining service life of gears according to claim 1, characterized in that, In step A2, the probability distribution of the experimental time-domain data is fitted based on a Gaussian mixture model, and the probability distribution of the experimental frequency-domain data after power transformation is estimated based on a Box-Cox mixture model.

3. The method for predicting the remaining service life of gears according to claim 1, characterized in that, The online phase includes: Step B1: Acquire the acceleration signal of the gear to be predicted and decompose it through a high-pass filter and a low-pass filter of multi-level discrete wavelet transform to obtain detail coefficients and approximate coefficients, and assemble them to obtain a coefficient sequence; Step B2: Divide the coefficient sequence into blocks and embed it to obtain an embedded coefficient sequence; Step B3: Extract local features and global features from the embedded coefficient sequence respectively, and fuse them to obtain fused signal features; Step B4: Predict the remaining lifetime based on the characteristics of the fused signal.

4. The method for predicting the remaining service life of gears according to claim 3, characterized in that, Step B2 includes: Step B21: Divide the coefficient sequence into multiple frame blocks according to a preset length; Step B22: Project the frame blocks onto a feature space of a specific dimension to obtain the embedding coefficient sequence.

5. The method for predicting the remaining service life of gears according to claim 3, characterized in that, Step B3 includes: Step B31: Extract wavelet inputs at different decomposition levels from the embedded coefficient sequence and input them into the LSTM model to obtain local features, and input the embedded coefficient sequence into the Transformer model to extract global features; Step B32: Fuse the local features and the global features to obtain the fused signal features.

6. The method for predicting the remaining service life of gears according to claim 5, characterized in that, Step B32 includes: The weights of the local features and the global features are calculated separately based on the self-attention mechanism and then fused.

7. The method for predicting the remaining service life of gears according to claim 5, characterized in that, Step B4 includes: Step B41: Based on the fused signal characteristics, a multilayer perceptron is used to perform nonlinear mapping to obtain the corresponding health indicators; Step B42: Determine the remaining lifespan based on the health indicators.

8. A storage medium, characterized in that, The method includes computer instructions, which, when executed by a computer device, perform the gear remaining service life prediction method as described in any one of claims 1-7.

Images